Decoupling Solutions

Integrated circuit (IC) technology has advanced over time, requiring innovative decoupling solutions that address decoupling needs over the entire power path from the main power supply to the logic gates on the IC. As one follows the path from the power supply to logic gate, high-frequency content increases, required capacitance falls, and the importance of very low inductance (ESL) increases. Impedance (Z) must also remain low and stable over the frequency range of interest so that detrimental resonances are avoided. Four tiers or levels of decoupling have been defined along the path from the power supply to the logic gates on the IC: (1) the voltage regulator level, (2) the board decoupling level, (3) the IC package decoupling level, and (4) the IC on-die decoupling level. Each level has impedance requirements that are best addressed by capacitor technologies which differ in capacitance density, capacitance stability, ESL, ESR and cost. This paper addresses the relative performance of competing capacitor technologies for board level (2) and IC package level (3) decoupling. New decoupling paradigms are presented for both of these levels. The Power Delivery Network (PDN) Power delivery to high-speed microprocessors cannot be accomplished using any single device or component. Large current changes (30 to 100 A) can occur during time durations from seconds to nanoseconds. Proper microprocessor operation requires the voltage remain stable during those current transients (typically less than ±0.1 V variation around a 1-V supply) The power supply alone cannot maintain that voltage stability at those current levels and the shorter time scales due to limited bandwidth. Decoupling capacitors are required in the power delivery network (PDN) between the power supply and the microprocessor to maintain voltage stability. μP VRM PDN Power Flow Cdie Cpackage Cmid Cbulk Response Time: 100’s ps 1’s ns 10’s ns 100’s ns 1’s μs Frequency: 1 GHz 100 MHz 10 MHz 1 MHz 100 kHz Stepped Load Power Source Figure 1. Progressively faster loops approaching the microprocessor (Load). Figure 1 shows a typical decoupling scheme used for microprocessors today. Several types of capacitors are required. These capacitors are used in progressively faster loops as the PDN approaches the microprocessor. The extreme right of the diagram shows the power supply or voltage regulator module (VRM). The best VRMs with bandwidths of 100 KHz are able to respond to current changes on timescales of μs to 10’s of μs. Moving to the left, the capacitors of Cbulk need to respond on time scale of 100’s of ns or at a frequency of up to 1 MHz. The next loop is handled by the capacitors of Cmid (for mid-frequency) that respond in 10’s of ns or up to 10 MHz. The capacitors for Cbulk and Cmid are located on the system motherboard. The final loop that can be chosen by the designer of the PDN is Cpackage (for the capacitors located on the microprocessor package). The response time needs to be in the nanosecond range or in the frequency domain up to 100 MHz. The final loop represents the capacitors on the microprocessor die (Cdie). Through many years of trial, the capacitor types chosen for Cbulk are aluminum or tantalum electrolytic capacitors and have a combined capacitance in the 1000’s of μF, an ESR of a few milliohms, and an inductance on the order of a nH. For Cmid, high-value MLCC capacitors with a combined capacitance of a few 100 μF, sub milliohm ESR, and inductance of 10’s of pH are chosen. For Cpackage, lowinductance MLCC capacitors with a combined capacitance of 10’s of μF and inductance on the order of a few pH are chosen. Within a loop, multiple capacitors are needed to satisfy requirements. 2008 ECA (Electronics Components, Assemblies & Materials Association), Arlington, VA Proceedings CARTS USA 2008, 28 Symposium for Passive Electronics, March, Newport Beach, CA The Cbulk and Cmid capacitors are normally found on the motherboard in close proximity to the processor. Because the inductance of Cpackage is so critical, these capacitors are mounted on the same small interposer PCB as the processor is mounted, above the pins or solder ball attachment, and as close to the silicon die as possible. Faster and more powerful microprocessors have resulted in increased current demands and faster response times. There is also a desire to decrease circuit size in order to decrease overall system size or to add more features to the same size system. This article discusses new decoupling schemes at the motherboard or package level that are responsive to these requirements. Theoretical Analysis of Motherboard-Level Decoupling Performance Simple analyses of PDNs have been accomplished by drawing circuits that contain multiple clusters of capacitance connected together by series resistance and inductance (see, for example, Figure 1). These clusters of capacitance are located between the microprocessor and the VRM. Such a circuit is useful for illustrating that each of these clusters of capacitance responds to a step change in processor current after its own, progressively longer time delay until finally the VRM takes over. Inherent in this simplified model is the idea that charge delivery by these capacitor clusters may not smoothly overlap, so there could easily be peaks and droops in the voltage response at the processor during the critical response time interval. Figure 2 illustrates this schematically. The ideal response to the step load in current is shown by the dashed line in the voltage plot. The effect of mismatched impedances causes non-uniformity in the voltage response, as shown by the solid line. Unfortunately, this analysis approach does little to quantify the expected voltage response in an intuitive way that mimics actual decoupling behavior. Thus it is not a very useful tool beyond communicating the general nature of the PDN and explaining why lower-ESL capacitors are located closer to the processor and higher-ESL capacitors are located farther away. Figure 2. Ideal and Non-Ideal Responses of a Microprocessor PDN. To fill this analysis gap, there is a powerful theoretical strategy for analyzing the decoupling performance of a PDN. The strategy is to characterize the impedance looking into the PDN from the microprocessor’s perspective and then calculate the time-domain step response of the voltage across this impedance to a step change in load current. Specifically, the frequency-domain complex impedance looking into the PDN from the processor’s point of view is either predicted from models of the capacitors and circuit board or measured on actual motherboards with an impedance analyzer. This frequency-domain complex impedance is then transformed into the time domain via inverse fast Fourier transform (IFFT) to discover the impulse response of the network. The impulse response is then integrated to discover the step response, and the step response is multiplied by the anticipated current step amplitude. The end product of this procedure is a time-domain plot of the theoretically-expected response of the PDN to a step change in microprocessor load current. If the impedance prediction or measurement is accurate, the resulting step response must also be accurate. 2008 ECA (Electronics Components, Assemblies & Materials Association), Arlington, VA Proceedings CARTS USA 2008, 28 Symposium for Passive Electronics, March, Newport Beach, CA After applying this analysis over a variety of cases, it quickly becomes clear that the best step response is found in PDNs that present the processor with the most constant impedance spectrum. Of course, most real-life PDNs are not perfect, and their non-constant impedance gives rise to non-ideal step response. This will be demonstrated presently, but first an example is given of a nearly ideal board-level decoupling solution.